When rate of netns creation/deletion is high enough, we observe softlockups in cleanup_net() caused by huge list of netns and way too many rcu_barrier() calls.
This patch series does some optimizations in kobject, and add batching to tunnels so that netns dismantles are less costly. IPv6 addrlabels also get a per netns list, and tcp_metrics also benefit from batch flushing. This gives me one order of magnitude gain. (~50 ms -> ~5 ms for one netns create/delete pair) Tested: for i in `seq 1 40` do (for j in `seq 1 100` ; do unshare -n /bin/true >/dev/null ; done) & done wait ; grep net_namespace /proc/slabinfo Before patch series : $ time ./add_del_unshare.sh net_namespace 116 258 5504 1 2 : tunables 8 4 0 : slabdata 116 258 0 real 3m24.910s user 0m0.747s sys 0m43.162s After : $ time ./add_del_unshare.sh net_namespace 135 291 5504 1 2 : tunables 8 4 0 : slabdata 135 291 0 real 0m22.117s user 0m0.728s sys 0m35.328s Eric Dumazet (7): kobject: add kobject_uevent_net_broadcast() kobject: copy env blob in one go kobject: factorize skb setup in kobject_uevent_net_broadcast() ipv6: addrlabel: per netns list tcp: batch tcp_net_metrics_exit ipv6: speedup ipv6 tunnels dismantle ipv4: speedup ipv6 tunnels dismantle include/net/ip_tunnels.h | 3 +- include/net/netns/ipv6.h | 5 +++ lib/kobject_uevent.c | 94 ++++++++++++++++++++++++++---------------------- net/ipv4/ip_gre.c | 22 +++++------- net/ipv4/ip_tunnel.c | 12 +++++-- net/ipv4/ip_vti.c | 7 ++-- net/ipv4/ipip.c | 7 ++-- net/ipv4/tcp_metrics.c | 14 +++++--- net/ipv6/addrlabel.c | 81 ++++++++++++++++------------------------- net/ipv6/ip6_gre.c | 8 +++-- net/ipv6/ip6_tunnel.c | 20 ++++++----- net/ipv6/ip6_vti.c | 23 +++++++----- net/ipv6/sit.c | 9 +++-- 13 files changed, 157 insertions(+), 148 deletions(-) -- 2.14.1.690.gbb1197296e-goog
